分类器动态组合及基于分类器组合的集成学习算法
Dynamic Combination Method of Classifiers and Ensemble Learning Algorithms Based on Classifiers Combination
作者：付忠良(1.中国科学院 成都计算机应用研究所，四川 成都 610041;2.中国科学院 研究生院，北京 100049)；赵向辉(1.中国科学院 成都计算机应用研究所，四川 成都 610041;2.中国科学院 研究生院，北京 100049)
Author：Fu Zhongliang(1. Chengdu Computer Applications Inst., Chinese Academy of Sciences, Chengdu 610041, China;2.Graduate Univ. of Chinese Academy of Sciences, Beijing 100049, China)；Zhao Xianghui(1. Chengdu Computer Applications Inst., Chinese Academy of Sciences, Chengdu 610041, China;2.Graduate Univ. of Chinese Academy of Sciences, Beijing 100049, China)
收稿日期：2010-07-27 年卷（期）页码：2011,43(2):58-65
期刊名称：工程科学与技术
Journal Name：Advanced Engineering Sciences
关键字：分类器动态组合；集成学习；多分类问题；AdaBoost
Key words：dynamic combination of classifiers;enable learning;multi-class classification problem;AdaBoost
基金项目：国家“863”计划资助项目(2008AAO1Z402)；四川省重点科技计划资助项目（2008SZ0100；2009SZ0214)
中文摘要
针对目前基于分类器静态组合的集成学习算法难于推广的问题，根据组合分类器分类错误最小化原则，研究了组合系数随分类器输出变化而变化的分类器动态组合理论,包括组合系数的选取、组合分类器分类错误率的估计等。证明了在分类器相互独立时，一些动态组合分类器等价于Bayes统计推断。提出了基于分类器组合的通用集成学习算法，并把AdaBoost、Real AdaBoost、Gentle AdaBoost算法推广到了多分类问题。证明了按照集成学习算法得到的分类器，其动态组合的有效性可不依赖于分类器的独立性，这支撑了基于分类器相互独立假设来研究分类器组合的有用性。最后,通过UCI数据实验验证了动态组合的有效性。
英文摘要
For the generalization problem of ensemble learning algorithms based on classifiers static combination, by minimizing the error of combination classifier, a dynamic combination method of classifiers was studied, in which the combination coefficients varied according to the output. Specially, the selection of combination coefficient and the estimation of classification error rate of combination classifier were researched. It was proved that some dynamic combination classifiers were equal to Bayes statistical deduction when the classifiers were independent of each other. The method for constructing general ensemble learning algorithm based on classifiers combination was put forward, and AdaBoost, Real AdaBoost, and Gentle AdaBoost algorithms were extended to solve the multi-class classification problem. It was proved that the efficiency of the dynamic combination in the classifier obtained by ensemble learning algorithm did not need the condition that the combined classifiers were independent. Therefore, the feasibility of classifier combination under the assumption that the combined classifiers were independent was obtained. At last, the dynamic combination method was verified efficient by the experiments on UCI dataset.
【关闭】
